Code Assist Beta: AI powered code completion

@ThatLazyBucket Code Assist won’t create the game or write an entire library with a single prompt (“–make this for me”) but it’s quite helpful at filling in the details (including writing entire functions) if you give it direction on the structure, as in the examples above. I hope this helps, thanks!

This looks extremely interesting. I am curious to see how it plays out once the beta is worked out, but for now its pretty glitchy. I would have to experiment with it more but I see amazing potential with this.

1 Like

At first I was not very happy with this, I thought the AI would just be generating code instead of making you actually have to know how to code (if that makes sense). I am super glad that it just helps finish or fix code instead of do it all for you.

2 Likes

It doesn’t seem to want to finish or see the previous code that I gave it. I tried a simple code that when the part is touched, it damages the player. I have the “touch” part added, but the AI doesn’t want to add the damage part.

That’s because you already have an onTouched function, and it assumes you could add the humanoid:TakeDamage to it.

The AI will not modify the previous code because of this reason. It’s coded not to do so.

2 Likes

So then how will it make me a part that damages you when touched? Should I add the damaging part first and then make it write the onTouched function?

I don’t think you know how to code. It doesn’t work like that. The damaging part goes inside the function, but since it is already written, the AI will not edit it.

1 Like


pretty lit if you know how to use it, but I would focus on studio.

This tool is amazing! However, I’ve noticed that the AI adds unnecessary ‘end’ keywords when i have already added those before making the prompt. Which can become annoying as I have to remove them everytime i use it.

image

image

2 Likes

The point is that AI will never be perfect (:sweat:) and will always need a human to intervene and refine the code.

3 Likes

Thanks for sharing your thoughts. This is a known issue. We are working on it.

1 Like

Copying chatgpt? hdtihgdfjiuy5dsddhyf

1 Like

This uses the api that chatgpt is using and it is called OpenAI. Over 500 websites are using it and it is rising more

1 Like

I genuinely wish for a utopia where you can make an AI print perfect results with just 10 tokens.
(10 words in a comment)

Sadly, we live on earth and primitive models need context. Finetuning takes you far but it’s not a human programmer.

It doesn’t even know what language you want without an existing script above it. Generative AI will repeat the best possible pattern it can. I got results in python when I tried a single comment prompt.

1 Like

ChatGPT uses OpenAI’s 2019 model.

From what I’ve heard, this AI uses the 2017 model and is about one order of magnitude smaller in size.

Although, no point guessing this stuff. Trust in the Roblox team!

It’s a glorified pattern repeater. It generates as if nothing is below it.

The AI can’t see below your script, only above I’m sure. So it thinks there’s no end’s below.

This is because the tokens Roblox feeds to it probably don’t include the underneath of your script for several reasons:

  1. Variables below the generated code can’t usually be used by the code it’s generating.
  2. These models expect their response to be the last thing in the string. They are an AI autocomplete.

If by this you mean something like:
“Create a fighting system with raycasting which fires when the A key is pressed” (I know it’s more than 10 but you get what I mean)
And it creates a perfect result, this would destroy little Programmers 100%
But I don’t want to restart arguing

One day… One day…

Pogrammer jobs will turn into prompt engineering jobs when that happens. That wont be a hard job to move to for us.

Mail marketing turned into Email marketing.
Software engineers will turn into prompt engineers.

1 Like

It seems to not be working for me I tried putting teams stuff nothing works.

Very poorly trained AI model.
Code that is generated sometimes provides out of date, depricated API methods.
Ask the community of verified developers, like myself, to train it instead of using the free model catalogue.
Invoking the AI when it refuses to generate code should provide a notice saying the code generation failed.
In it’s current state it is not only unusable, but also harmful, as it can provide suggestions to new developers that don’t understand code, even when the code is flawed, and depricated.
I have given this feature a few weeks for users to report bugs, including repetition, which has not been fixed. I can confidently recreate repetition conditions using numerous methods.
This feature needs to be entirely reworked.
Many times when I have tried to use this AI, I’ve gotten bad code.
I am willing to volunteer my work to train the AI using new Luau methods, such as type checking and threaded workloads. None of my “free models” are updated using any new coding methods. Any free models that I see are almost always not going to be using proper methods, especially older models to which utilize almost all depricated events, or using coding workarounds to bypass the limitations of Roblox that existed prior to the creation of newer methods.

2 Likes